Contents

NVIDIA has added five models to Nemotron Retriever, its suite for making internal documents and other data easier to search and use with AI. The new models handle embeddings, search result reranking, and PDF parsing, with options for different use cases and scales.

Key Points#

  • The new releases include three embedding models, one reranking model, and one PDF parsing model.
  • The 8B embedding model supports 34 languages, including Japanese, and handles both images and text.
  • NVIDIA’s comparisons show its embedding and PDF parsing models outperforming competing products in accuracy and speed.
  • All five are available as open models under the OpenMDW-1.1 license.

What the Five Models Do#

RAG converts source materials into searchable data, finds results relevant to a question, and passes them to generative AI. These models support the process from preparing data for search to reranking results.

The embedding model, Nemotron-3.5-Embed-8B-BF16, converts text and images into representations usable for search. It supports 34 languages, including Japanese. NVIDIA also released the smaller Nemotron-3.5-Embed-1B-BF16 and its NVFP4-quantized version, Nemotron-3.5-Embed-1B-NVFP4.

Nemotron-3.5-Rerank-8B-BF16 evaluates how relevant images and text are to a question and ranks search results. Nemotron-3.5-Parse converts PDF content into data that AI can process more easily. This setup assigns search data preparation and result selection to separate models.

Benchmark Highlights#

In comparisons published by NVIDIA, the 8B embedding model scored higher than Qwen3-VL-Embedding-8B and EmbeddingGemma-2. The company says the 1B BF16 version also performed close to Qwen3-VL-Embedding-8B. On a chart comparing processing speed and accuracy, the 8B, 1B, and 1B NVFP4 versions are positioned favorably against competing models.

The company also explains that combining an embedding model with Rerank can improve accuracy while keeping per-query costs down. For Parse, it published a chart comparing pages processed per second and accuracy, highlighting its higher processing speed relative to competing models.

These evaluation results come from NVIDIA. Publicly available information alone cannot establish whether your documents and questions will yield the same accuracy, speed, and cost. Test with real data before deciding to adopt the models.

What to Check When Testing#

First, check the types and languages of your source materials. If they are mainly PDFs, include Parse; if you also need to search images, consider an Embed model that supports image input. If computing resources are limited, start with a 1B model and compare it with the 8B model as needed. The quantized version is another option, but verify its impact on speed and accuracy in your actual environment.

For evaluation, prepare a few dozen representative PDFs and about 10 questions users are likely to ask. Along with search result accuracy, record document ingestion speed, how often incorrect results appear near the top, and per-query costs to make operational differences easier to assess.

Public availability does not mean every use is permitted under the same conditions. Check the OpenMDW-1.1 terms on the distribution page against your intended use and redistribution method.

Next Step#

Choose a few dozen representative PDFs, build a small search dataset with Embed-1B and Parse, and check the results for 10 common internal questions.

References#

Source: https://news.livedoor.com/article/detail/32628832/

Contents